By Guest 
Undergraduate students with limited AI expertise claim to have built a new AI model capable of generating podcast-style content similar to Google’s NotebookLM.
Students Build AI Competitor to Google’s NotebookLM
Two undergraduate students with no formal AI training claim to have developed a new artificial intelligence model that can generate podcast-style content similar to Google’s NotebookLM. According to TechCrunch, Toby Kim, co-founder of Nari Labs, said he and a colleague began learning about speech AI models just three months ago. Inspired by NotebookLM, they aimed to create a model that gives users more control over generated voices and creative scenarios.
Introducing Dia: The Customizable Voice Model
The team used Google’s TPU Research Cloud to train their model, named Dia, which offers free access to Google’s AI chips for researchers. Key features include: – 1.6 billion parameters for script-based dialogue generation – Customizable speaker tone and style – Ability to add non-verbal cues like coughs, laughter, and pauses – Random voice generation when no specific instructions are given – Potential for voice imitation (with ethical concerns)
Availability and Limitations
Dia is currently available on Hugging Face and GitHub and can run on most modern PCs with at least 10GB VRAM. However, the model has: – Minimal safeguards against misuse – No disclosed training data (potentially including copyrighted material) – No liability assumed by Nari Labs for misuse
This student project demonstrates how accessible AI development has become, while raising important questions about ethics and copyright in open-source AI models.